alphaedge-ai/NanoBEIR-clean

NanoBEIR clean

Cleaned version of lightonai/nanobeir-multilingual (en and fr splits) in the sentence-transformers / MTEB format (split into corpus / queries / qrels), with one configuration per language.

Each row contains a subset column that allows you to filter for the desired subset.

Data Cleaning Applied

Issues reported in mteb#3708:

• empty or excessively short documents (LENGTH(text) < 5, < 3 words);
• LLM translation meta-prompts left in the documents (fr);
• data leak with SFT LightOn data (embeddings-fine-tuning-filtered-{en,fr}), removed by normalizing the text (accents folded);
• duplicate text in the corpus.

Usage

from datasets import load_dataset

datasets = [
    "NanoArguAna",
    "NanoClimateFEVER",
    "NanoDBPedia",
    "NanoFEVER",
    "NanoFiQA2018",
    "NanoHotpotQA",
    "NanoMSMARCO",
    "NanoNFCorpus",
    "NanoNQ",
    "NanoQuoraRetrieval",
    "NanoSCIDOCS",
    "NanoSciFact",
    "NanoTouche2020",
]

language = "fr"   # "en" or "fr"
repo = "alphaedge-ai/NanoBEIR-clean"

corpus  = load_dataset(repo, f"NanoQuoraRetrieval_{language}", split="corpus")
queries = load_dataset(repo, f"NanoQuoraRetrieval_{language}", split="queries")
qrels   = load_dataset(repo, f"NanoQuoraRetrieval_{language}", split="qrels")